PCSAFace: Face Recognition with Parallel Channel and Spatial Attention
Haitao He, Weimin Zhou, Xiangqian Liu · 2021
Recently, face recognition task has achieved a tremendous strides. It's widely accepted that the key to discriminate different identities is to enlarging the distance between different identities and decreasing the distance between different faces of same identity. Most researches focus on decreasing the intra-class distance and achieve a nice result. There is a trend that attention module is introduced into computer vision task. In this paper, we propose the "angular regularization" that focuses on enlarging the inter-class distance and introduce an novelty parallel channel and spatial attention block to build a new face recognition dedicated backbone network. The method, we call it PCSAFace, enlarges distance between different identities by directly penalizing the angle between cluster center of one identity and cluster centers of other identities. PCSAFace has a well defined geometric interpretation and is easy to do more in-depth research in face tasks. The method has no less good performance than previous works.